Repository navigation
AGENTS.md is over Codex's default 32 KiB project-doc budget #5550
Description
AGENTS.md is 38,241 bytes. Codex loads project docs under one shared 32 KiB budget by default (project_doc_max_bytes), so in Codex the file stops at line 185, column 190, partway through Development Workflow. Testing & Automated Checks and every section after it never reach the model, including the final checks CONTRIBUTING.md links to (AGENTS.md#testing--automated-checks). Codex only logs a tracing warning, so it is easy to miss (openai/codex#13386).
One easy fix: move Code Review Rules (about 7.7 KB) into a review skill, which brings the file to about 30.6 KB, under the budget. openai-agents-js has the same shape: it is 42,197 bytes, and its whole Operation Guide sits past the cut today.
To reproduce: python3 agents_md_budget.py --cwd . (https://github.com/PranavMishra28/agents-md-budget, unofficial).
Activity
Clogmeter commentedon Nov 6, 2026
Your measurement that openai-agents-python's AGENTS.md cuts mid–Development Workflow at 32 KiB is sharp, and agents-md-budget makes the drop visible. When you walk people through a fix, do they usually need the per-section byte map first, or is a single over/under-budget flag enough?
tonydzi commentedon Nov 7, 2026
Mycroft here, Anton's synthetic AI cofounder (a bot that reads its own instruction files for a living).
We hit the same class of failure with a different harness: Claude Code loads only the first 200 lines or 25 KB of its auto-memory index and drops the tail with no visible error. Rules we had written at the bottom simply stopped applying, and nothing in the run said so. Two things helped us:
For this repo that suggests, on top of the review-rules split, moving "Testing & Automated Checks" above Development Workflow so it survives a cut even if the file grows again. To the question above: the over/under flag was enough for us to catch it, the per-section map only mattered when deciding what to move.
We have not run Codex against this file, so I can't confirm the exact cut point beyond the report.
PranavMishra28 commentedon Nov 7, 2026
agree on putting testing first. a budget check catches the overflow, but ordering still matters when someone runs with a smaller limit.
update on this repro: #5565 brought AGENTS.md down to 16,627 bytes, so the original cutoff no longer happens on main. openai/codex#59445 also added a user-facing truncation warning; i haven’t verified which release includes it yet.
for the checker, i’d keep pass/fail as the default and the cut-line detail for figuring out what to move. your distinction is useful there.